Answers
Shared memory for law firms: one position, every partner's AI
A firm's position on a question belongs to the firm, not to whoever was asked first — but the AI each partner uses knows only what that partner told it. Shared memory closes the gap at the source: the position is written once, a person at the firm sanctions it, and from then on it is served to any AI any partner connects, before that AI starts work. Nobody re-explains the house view in a prompt, and nobody discovers that a colleague's answer contradicted it after it has already gone to a client.
Last updated September 1, 2026
The position belongs to the firm; the prompt belongs to a person
Every partner who works with an AI builds a private version of the firm inside it — how a caveat is phrased, which side the house takes on a recurring question, which clause never gets signed off. None of that is visible to anyone else, and none of it survives the partner changing tools. What the firm has decided is the one thing that cannot live there, because it has to reach the work of people who were not in the room when it was decided.
Sanctioned once, read by every AI at the start of the work
A position is written as a rule, and a person at the firm approves it before it governs anything — an assistant may propose, never promote its own inference. From that moment the rule travels on the read that opens an AI's work, ranked against the matter at hand, whichever assistant that partner uses. Arroway carries the norm; it carries no opinion about which tool anyone should be working in.
Every position carries the thing that ends it
A firm's position changes: a court rules, a statute is amended, the house revises its own reading. Each memory is stored with the condition that kills it or forces a review — an explicit revocation, a named trigger, a date. A superseded position stops being served instead of quietly competing with the current one, and the record of what it used to be stays for whoever asks why the firm answered differently last year.
What it looks like in practice
Two partners are asked the same question about a limitation-of-liability clause, by clients who will never meet. One answers that the firm accepts it with a carve-out; the other, that the firm does not accept it at all. Both answered from their own AI and both answered in good faith — the house settled this two years ago, in a matter neither of them worked on. The decision existed. It simply had no way of reaching either of them.
Questions people ask about this
- Does this mean the AI decides the firm's position?
- No, and the product is built around that. An assistant writes and may propose; only a person turns a proposal into a rule in force. What an assistant inferred and what the firm decided are stored as different things and served differently, so nobody reading later can present one as the other.
- Our partners use different AI tools. Does everyone have to switch?
- No. The memory lives outside the assistants and is read by whichever one each person has connected, so a firm where one partner works in one tool and another in a different one still keeps a single position between them. Changing tools does not carry the firm's decisions away with it.
- Who at the firm has to maintain this?
- Less upkeep than a shared document would ask for, because the writing happens as a side effect of the work: the assistants write, and a person approves or corrects in a short review after the fact, never as a gate the work waits behind. Curating the collection is the product's job, not a standing chore for whoever owns the account.
Where this is verifiable
Product documentation on this site (How it works, Install), the sanctioned product spec in the repository, and the Privacy Policy for what is stored and who can reach it. Everything described here is behaviour the tools apply today, not roadmap. This page makes no claim about the built-in memory of any AI product, and it is not legal advice about professional-conduct duties in any jurisdiction.
https://www.arroway.app/en/answers/shared-memory-for-law-firms